Agentic performance is largely invariant to work interruptions. Human performance is not.

This sounds obvious but the implication is underappreciated. It is not just that an agent is faster or cheaper than a human at some task. It is that the task was literally not completable given the human’s available schedule. Twelve ten-minute intervals over three days is not a viable way for a human to hold and advance a complex disentangling problem. Context evaporates between sessions. But twelve feedback iterations from a human steering an agent — even a poorly-performing one — can actually move the work.

The relevant class of work here is deterministic and “boring” but highly complex: unpicking someone else’s plumbing, porting a pattern from one codebase to another, reverse-engineering a reference implementation. Pre-agent, this required long uninterrupted stretches. The cognitive load of re-establishing context from scratch each session was high enough that humans rationally avoided the work or deferred it indefinitely.

Scattered human time, previously nearly worthless for this class of problem, is now productive.